anthropics/healthcare: The Repo That Turns Claude Into a Resumable Workflow Engine

Inside the waypoint system, MCP servers, and deterministic checks that make healthcare AI feel less like chat and more like infrastructure.

9 min read • View on GitHub • More from anthropics

By Repo Explainer • AI-assisted • Published by Ron Kurti

Published Apr 8, 2026 • Coding Agents

A long archival corridor with filing cabinets and nested drawers, each drawer labeled as a stage in a healthcare workflow. It shows the repository treating Claude like a stateful operator that moves through checkpoints instead of improvising in one long prompt. — The repo is built around stages, not turns. Each stage saves state so the next one can pick up where it left off.

Key Takeaways

The repo treats Claude as a resumable workflow engine, not a single-turn chatbot.
Waypoints and deferred loading keep long healthcare tasks stateful, smaller, and easier to audit.
MCP servers move healthcare work from model memory to live, source-backed retrieval.
Deterministic Python code handles math the model should not guess.

The real product is not a chatbot

anthropics/healthcare is interesting because it refuses to behave like a giant prompt. The repo turns Claude Code and Claude Desktop into a staged operator that writes state to waypoints, loads only the next instruction set, and hands the parts that need certainty to code and live data.

The waypoint model turns one long healthcare task into a chain of saved stages. Each step leaves behind an artifact the next step can trust.

Why waypoints beat one giant prompt

The clever part is not the folder. It is the discipline around loading. The repo tells the model not to pre-read subskill files, then brings each one in only when the current waypoint needs it.

A tight view of folders, stamps, sealed envelopes, and a thread linking one checkpoint to the next. It explains deferred loading, where the next instruction set stays sealed until the workflow reaches the correct waypoint. — Deferred loading keeps the instruction set narrow. The workflow opens only what it needs, when it needs it.

That matters because healthcare work punishes improvisation. A protocol draft, an authorization appeal, or a registry lookup can span multiple passes and multiple sources. If the model is forced to hold all of that in context, every turn becomes a memory test. If the repo writes the state down, the system gets resumability instead of vibes.

How the repo grounds the model in real data

The MCP layer is the other half of the design. PubMed, the NPI Registry, and the CMS Coverage Database are not decorative integrations. They move the system from recall to retrieval, so Claude can assemble an answer from live sources instead of leaning on whatever was nearby in training data.

Source	Used for	Why it matters
PubMed	Research retrieval and literature review	Keeps clinical claims tied to citable evidence
NPI Registry	Provider identity lookup	Reduces the chance of building workflows on the wrong clinician or organization
CMS Coverage Database	Coverage and policy checks	Anchors administrative work in the rules that actually govern it

clinicians often spend more time on documentation than on patient care.

Mike Krieger, Chief Product Officer at Anthropic · UBOS on Anthropic

A desk with three conduits feeding into a protocol page from a medical library, a provider registry, and a coverage database. It shows the repo assembling healthcare output from live sources instead of model memory. — The model is not expected to remember every fact. It is expected to fetch the right ones and assemble them well.

The deterministic backstop: math the model should not guess

The repo draws a hard line around arithmetic. Its sample_size_calculator.py uses Python, SciPy, and NumPy for continuous sample-size calculations, which is exactly the kind of work you do not want a model improvising. The LLM can orchestrate the request, but the calculation itself should be deterministic.

A workbench with a calculator, ruler, and protocol sheet, plus a small mechanical device feeding numbers into the page. It explains why sample-size math is handed to Python and SciPy instead of asking the model to improvise. — When the answer has to be numerically defensible, the repo hands the job to code.

LLM guess	Python and SciPy	Result
Estimates numbers from context	Computes alpha, beta, dropout, and allocation ratio	A reproducible sample-size answer
Can sound confident while being wrong	Produces a deterministic calculation	A result reviewers can trace
Treats math like language	Treats math like math	A safer backstop for regulated work

What the healthcare workflows actually optimize

Once the architecture clicks, the use cases stop looking random. They all share the same shape: structured inputs, source checks, repeatable outputs. That is why this repo is pointed at work that is expensive to do manually and painful to get wrong.

Clinical trial protocols need source-backed steps and reproducible calculations.
Prior authorization workflows need policy lookups and durable evidence trails.
Claims appeals need structured reasoning that can survive review.
FHIR-related work needs a clear map from messy data to interoperable output.

How this differs from generic AI assistants

Split scene with a messy chat window on one side and a structured, file-backed workflow on the other. It contrasts generic assistant behavior with the repo's checkpointed process design. — Generic chat optimizes for a fast answer. This repo optimizes for a process that can be resumed, checked, and reused.

Generic chat-first assistant	anthropics/healthcare	Why it matters
Keeps state in conversation context	Writes state to waypoint files	Saved state survives long tasks and handoffs
Answers from model memory	Queries PubMed, NPI Registry, and CMS Coverage	Live sources reduce hallucination risk
Tries to solve everything in one pass	Breaks work into staged artifacts	Stage boundaries make review and recovery easier
Lets the model estimate math	Delegates calculations to Python	Deterministic numbers are safer in regulated work
Has one final answer	Leaves an audit trail at each waypoint	Reviewers can inspect the process, not just the output
General purpose	Built around healthcare workflows	Domain-specific structure beats generic flexibility

What to take away from the architecture

The useful lesson is not that healthcare needs more AI hype. It is that operational work needs explicit process. Checkpoints, selective loading, live sources, and deterministic backstops are what let a language model behave like infrastructure instead of autocomplete.